Bug#868861: apache2: Package upgrade does not play well with multiple instances and restarting them
I use multiple apache2 instances facility on several of my servers using the
debian example script to set them up - i.e. they're set up using
When doing security updates via apt-get upgrade then the postinst
script doesn't restart anything other than the default instance.
This obviously isn't much of a surprise (although it would be nice
if there was a way of registering them for restart on upgrade).
However it strikes me that this behaviour might be something worth
noting in the README.multiple-instances file though as it means the
admin will need to notice when apache2 is upgraded and manually restart
the other instances.
That's not the main reason for this report though - the main reason is
that after the postinst script has restarted the main instance if you
then try manually restart the other instances via any mechanism which
uses /etc/init.d/apache2 and if the /usr/sbin/apache2 binary has been
updated by the upgrade then you get:
[FAIL] Restarting web server: apache2 failed!
[....] There are processes named 'apache2' running which do not match your pid file which are left untouched in the name of safety,|Please review the situation by hand. |
The reason for this appears to be down to the behaviour of pidof which
the /etc/init.d/apache2 script uses to perform this safety check.
It appears that debian's pidof will first try find process matches by device/inode
pair (i.e. where the process executable's device/inode pair matches the device/inode
pair of the binary, if it's located on a local filesystem). Only if it finds no
matches that way does it then try to find matches based on full path name instead.
Straight after upgrading apache the binary /usr/sbin/apache2
has a different device/inode pair to that of all the running /usr/sbin/apache
processes (as they were run from the prior binary version which is now deleted).
So when the postinst script for apache triggers a restart of the main instance
then pidof in the safety check will list pids of the all instances (because it finds none
matching the exact device/incode pair, so falls back on finding ones matching
the path). Thus that restart will pass the check and proceed ok.
After that restart the main apache2 instance is running from the new binary
and so its device/inode pair matches /usr/sbin/apache. Running pidof
/usr/sbin/apache2 will now thus only list the pids of the main apache instance
processes because it finds at least one matching the device/inode pair in
a running process and so it doesn't list those processes from the other instances
(because they're still running off the old now-deleted binary with a different
device/inode pair). Hence when you try use the init script for the instance to
stop/restart apache you now get the error about not matching the pidfile.
E.g. these will error while the system is in that state:
service apache2-instancename restart
The apache2ctl command has no such check on the contents of the pidfile I believe
and so will still work, so e.g. this will work:
And also you can manually kill the old process using the pid in the pidfile and then
use service start or the init.d script for the instance.
Once the other instances running off the previous binary have been restarted using either
of these interventions then pidof will start listing them again.
I've no idea how you might resolve the issue though other than maybe weakening
the pidof check in the init script to just check for "apache2" instead
of the full path of "/usr/sbin/apache2" or alternatively just document
this issue in the README so at least people using multiple-instances know that upgrading the
apache2 package need to watch out for this (and perhaps just stop all
instances prior to doing the package upgrade so they don't have any
running off the prior binary).
I've had a quick look at the apache2 source in experimental and it seems to have
the same check in the apache2 init script, so this report should be valid against
the latest version.